In general, non-symmetric clustered file system is the file system which a metadata server (MDS) that manages the metadata of files, multiple data server (DS) that manage data of a file, and multiple client systems that store and access a file are connected on Ethernet network and operated as a communication protocol.
Multiple data servers are provided as a single storage, and can be managed as a free storage space by adding/deleting a data server or a volume in a data server. Thus, in the system managing multiple data servers, considering the failure rate that is proportional to the number of servers, a method distributing data like RAID level 5 and preparing a parity for a recovery or a mirroring method preparing a replica of the data, etc can be used. Here, while mirroring technology has a low efficiency due to duplication of data, a data distributed storage structure using a parity has a relatively high efficiency and can meet fault tolerance and so it is preferable.
A data distributed storage structure using a parity ensures the availability of data if a data failure occurs. For example, configuring a stripe with multiple parity chunks and parities generated from them, even if a failure occurs in the data chunk at the same time by the number of parities generated, a recovery is possible without losing data. On the other hand, in the prior, a partial parity generated before updating the data is managed in a local storage such as a data chunk. Thus, the conventional method is efficient in write performance aspect by reducing the time calculating and storing a partial parity. However, if a failure occurs in the data server which a partial parity chunk is stored before update of a parity, the updated part of the data chunk can be might difficult to recover. Also, in degraded read or degraded write processing for failure data, first, to check for updates of parity, check a partial parity chunk from a remote server, if necessary, since it is necessary to read a partial parity from the remote server and to process the parity update, a overhead on the network occurs, the performance of degraded mode also deteriorate. Therefore, conventional methods, there is a problem that the performance degradation of the input and output of data is sustained until the restoration of failure.